epfml / dynamic-sparse-flash-attention

Dynamic Sparse FlashAttention

Code to reproduce results for the paper "Faster Causal Attention Over Large Sequences Through Sparse Flash Attention"

Setup

To install the required python dependencies, first run:

pip install -r ./requirements.txt

Then install Triton:

git clone https://github.com/openai/triton.git
cd triton 
git checkout b2a757d00028fe844a93904036a18e8670bfe92f
cd python 
pip install cmake
pip install -e .

In the command above we set the Triton library to the commit used in our experiments. Feel free to experiment with later Triton versions.

Reproducing our LM experiments on OpenWebText2

GPU requirements: Preferably, you need at least one A100. Some of our experiments use data-parallelism with up to 3 A100s. You should have no problem running those experiments on any GPU supporting bfloat16, you might have to change the model parameters to adapt to the memory available.

Go in the openwebtext2-experiments folder and run the script/train-LMs.sh command.

Reproducing our runtime results

GPU requirements: We used one A100.

For the Hash-sparse and QK-sparse results, go in the runtime-experiments folder and check the timeperf-hash-and-qk-sparse.ipynb notebook.

Reproducing our Reformer results

Coming soon